NCCL Installation & Environment

This guide assumes you will run NCCL examples on Linux with NVIDIA GPUs.

Minimum environment checklist

• NVIDIA GPU(s)
• NVIDIA driver installed and working (nvidia-smi)
• CUDA toolkit available (or at least CUDA runtime)
• NCCL library installed
• A launcher for multi-process runs:
  • MPI (e.g., OpenMPI) is the simplest for examples
  • or SLURM / your cluster launcher

Typical cluster setup notes (conceptual)

• One process per GPU is common for examples
• NCCL will choose transports based on what is available:
  • NVLink / NVSwitch inside node
  • PCIe paths to NICs
  • Infiniband / RoCE / TCP between nodes

Helpful environment variables (starter set)

Use these when debugging:

• NCCL_DEBUG=INFO (or WARN)
• NCCL_DEBUG_SUBSYS=INIT,NET,GRAPH,COLL
• NCCL_IB_DISABLE=1 (force non-IB network path for A/B tests)
• NCCL_P2P_DISABLE=1 (disable P2P to isolate topology issues)
• CUDA_VISIBLE_DEVICES=... (control mapping)

The exact set you use depends on your fabric (IB vs RoCE vs TCP) and how your cluster is configured.

Next: Communicators